Tomasz Tunguz, the venture investor writing on X, picked up a disclosure AI Insiders covered this week, in which OpenAI said its research staff now gets more than three units of machine effort for each unit a person puts in during a normal workday. His post argues that number describes a clock, not a skill: a researcher gets credited with a multiple because software keeps working after the office empties out, not because that researcher thinks any faster.
The distinction matters for how the figure gets used in hiring and budget conversations. A true productivity multiplier should scale with headcount, since a sharper engineer stays sharper no matter how many colleagues join the team. Wall-clock parallelism does not behave that way. Adding a second researcher buys a second set of overnight agent runs and a second matching inference bill, not a second multiple of insight. Judged by that standard, the honest comparison is dollars spent per finished, unattended task rather than a ratio of machine hours to a shift, because the ratio counts a run still being cleaned up the next morning the same as one that shipped.
Tunguz backs the point with spend figures he attributes to OpenAI’s own disclosure. He writes that the typical researcher’s daily inference cost sat near $14 in late March and had passed $600 by the middle of August, a jump he puts at roughly forty times in under half a year. He adds that researchers in the top decile of usage were billing above $7,000 a day, which he annualizes to about $2.5 million per person. Those are Tunguz’s readings of one company’s disclosed medians and a single upper percentile, not independently audited figures, and they describe cost, not output.
His structural argument is that this spending looks nothing like a capital purchase. A factory buys a welding robot once and then has every incentive to keep it running around the clock, because the machine depreciates whether it is idle or busy. A researcher renting inference by the token owns no such asset. Every extra hour of agent runtime is a fresh charge, decided in the moment rather than amortized by a finance team. That, in Tunguz’s account, is what lets an individual choose to keep several agents going overnight: no depreciation schedule forces the decision, only the bill that follows it.
He also notes, again citing OpenAI’s own numbers, that more than half of the multi-hour agent tasks completed in recent months still needed a human to fix them afterward, which is why he expects visible output to lag the raw ratio rather than track it one for one.
For any team pricing an agent rollout against headcount this quarter, the number worth tracking is not agent-hours per employee. It is the dollar cost of a task that ships without a human cleanup pass, benchmarked against what that task costs a person to do directly, since that is the figure that will show up in next year’s budget review.
Tomasz Tunguz laid out this reading in a post on X on September 8, 2026.
